Gross margin floors in AI-native B2B SaaS: how we underwrite the cost stack.

We have written before, in AI doesn't lower SaaS prices, it widens margins, about the long-term direction of travel — AI-native businesses operated for the long run should structurally earn more margin than their pre-AI predecessors, not less. The essay below is the underwriting half of that thesis. The long-term margin thesis only survives if the business has a defensible floor today and a defensible glide path tomorrow. A lot of businesses do not, and the diligence work is in separating the two.

Why the headline margin number is misleading

Traditional B2B SaaS has a single dominant cost line — hosting and a small allocation of customer support. Steady-state gross margins of 78% to 85% are the norm and the variance between businesses is small. AI-native SaaS does not behave that way. The cost line is bimodal. There is the legacy hosting cost, and there is the model inference cost, and the inference cost can be anywhere between five and forty percent of revenue depending on the workflow, the model choice, the prompt design, and whether the buyer is on a per-seat plan or a usage plan. The reported gross margin on the last twelve months tells us very little until we have decomposed the cost stack.

The three axes that determine the floor

We look at three axes when we are trying to establish where the margin floor actually sits.

1. Inference cost share

What share of revenue is currently being paid out to model providers, and how has that share trended over the last eight quarters. A business whose inference cost share is 22% today and was 28% a year ago is on a healthy glide path — the team is optimizing prompts, caching results, and substituting smaller models where appropriate, and the cost line is following the broader market down. A business whose inference cost share is rising over the same period is in trouble, and the trouble is usually that the product is doing more work per dollar of revenue each quarter without a corresponding pricing response.

2. Model dependency

How dependent is the product on a single model from a single provider, and how easily could the team substitute. Businesses that have built around a single closed model with no abstraction layer in between will, eventually, eat a margin shock when that provider changes pricing, deprecates a model, or restricts a usage pattern. Businesses that have invested in a model-agnostic layer and routinely route different parts of the workflow to different models can substitute and survive. The cost of building that abstraction layer is real and we expect to see it on the engineering roadmap.

3. Pricing model alignment

Is the revenue model aligned with the cost model. A per-seat product with a heavy usage-based cost line is a structural margin problem the operator has to solve. A usage-priced or outcome-priced product whose pricing scales with the same units the cost scales with is structurally healthier. The intermediate case is a tiered or hybrid pricing model that caps usage inside the plan and meters overage. The pricing model does not have to be pure usage to clear the bar — it has to be defensible against the cost shape.

Margin in AI-native SaaS is a system design choice, not a line item. Businesses that did not design for it almost never retrofit it.

The Cobalt Glacier floor

Our underwriting bar for an AI-native acquisition is a defensible steady-state gross margin of 70%, with a clear and credible path to 80%+ over a five-year window. The 70% is not arbitrary. It is what we believe is required to support the free cash flow conversion bar we wrote about in Free cash flow conversion is the underwriting bar. A business that runs at 60% gross margins can still be a real business, but it does not produce the cash conversion profile a permanent-capital holding period needs to fund its own compounding without external capital. We are uninterested in owning a business whose growth is structurally tethered to outside funding.

The glide path to 80%+ is what makes the 70% floor more than a single-period filter. We expect three forces to lift the margin over the holding period — falling inference costs across the market, internal optimization of the product's model usage, and the expansion of the product into adjacent surface area that carries traditional software economics rather than inference economics. The 80% is achievable but not guaranteed. The discipline is in underwriting to the 70% floor and being delighted when the glide path delivers above it.

Diligence on the floor

The cost decomposition is not a single slide. It is a workstream. The work we actually run in diligence looks like this.

• Cost-of-revenue teardown. Hosting, inference, third-party tooling, payment processing, customer support, and implementation services are each pulled out as separate lines with eight quarters of history. We want to see each line as a percentage of revenue, not as an absolute number.
• Unit-economics by workflow. Where the product has more than one major workflow, we want unit cost and unit revenue for each. Margin blends are usually hiding a high-margin workflow that is funding a low-margin one. The low-margin one is the underwriting question.
• Model substitution simulation. The engineering team walks us through how the product would behave if it had to substitute the current primary model with the next best alternative — both upmarket and downmarket. The exercise surfaces hidden dependencies and the team's actual operating discipline around them.
• Pricing repricing model. We model the gross margin under three pricing scenarios — current pricing, a packaging refresh inside the existing model, and a structural move toward usage-aligned pricing. The output tells us how much of the margin gap we have to fix through engineering and how much we can fix through commercial work.

What disqualifies a business

Several patterns will end the conversation regardless of the topline growth.

• Rising inference cost share with no pricing response. The trend line has to bend at some point. If it has not bent in eight quarters, we do not believe it will bend after close.
• Single-model dependency with no abstraction layer. We are not interested in being a single provider's distribution channel on margin terms we do not control.
• Pricing model fundamentally misaligned with cost shape. A per-seat product with usage-shaped costs is a margin structure that has to be re-architected, and the re-architecture is hard to do without disrupting the customer base. We will rarely take that work on inside an acquisition.
• Operator who cannot describe the floor. If the founder cannot articulate where their margin floor sits and why, they are not operating the business with the cost discipline a permanent-capital hold requires. That is a coachable gap in a partnership, not a problem we will buy ourselves into solving from cold.

The bottom line

Gross margin in AI-native B2B SaaS is a designed outcome, not a reported one. Cobalt Glacier underwrites to a 70% steady-state floor with a credible path to 80%+, decomposes the cost stack line by line, and walks away from businesses whose margin structure depends on a single provider, a misaligned pricing model, or an operator who has not done the work. The compounding thesis of AI-native software only survives if the floor is real. The discipline is in checking the floor before the demo wins the room.

If you are a founder running an AI-native B2B SaaS business and want to talk through the margin floor with us before any diligence process, start a conversation. If you are an LP or investor thinking about how the next decade of AI-native economics will compound, our investor page explains the broader frame.